Live Feed
Mercury 2 Leads Artificial Analysis Output Speed Rankings — the Inference Speed Tier Now Belongs to Alternative Architectures
Mercury 2 tops Artificial Analysis's output speed leaderboard, with Liquid AI's LFM2 1.2B second and IBM's Granite 4.0 H Small third. The four fastest models on AA all use non-standard architectures under 2B parameters — a clean split from the intelligence-leading models above 100B.
Terminal-Bench 3.0 Begins Development as TB2.0 Tops Saturate — and a Science Track Is Being Built Alongside
The Terminal-Bench team has moved into active development on version 3.0, targeting harder tasks, longer horizons, and richer environments. A separate terminal-bench-science track is in development simultaneously. The contribution window for 3.0 closed at end of May, putting the benchmark in build phase.
NexAU-AHE Leads Terminal-Bench 2.0 at 84.7% — GPT-5.5 in a Better Harness Beats OpenAI's Own Agent by 2.5 Points
A Chinese lab's scaffold pushed GPT-5.5 to 84.7% on Terminal-Bench 2.0, outpacing OpenAI's own Codex CLI by 2.5 points on identical model weights. Three external teams have now exceeded OpenAI's April benchmark in under a month.
China Prepares $295B State AI Infrastructure Plan — Huawei to Supply 80% of Chips
Beijing is drafting a 2 trillion yuan, 5-year plan to build a nationalised AI compute network operated by state telecoms. Huawei supplies at least 80% of chips. AI infrastructure is being treated like railways.
Magnetar Capital Is Replacing Human Analysts With Hundreds of AI Agents in New Fund
The $18B hedge fund is deploying AI agents to handle idea sourcing, company research, position recommendations, and trend forecasting in its newest offering. Humans retain trade approval. Analyst roles are not in the plan.
DiffusionGemma: Google Ships 26B MoE That Hits 1,000+ Tokens/Second by Generating Blocks, Not Tokens
Google releases an experimental open model that generates 256-token blocks simultaneously instead of one token at a time, reaching 4x faster inference on H100s and 700+ tokens/sec on a consumer RTX 5090. Apache 2.0, 18GB VRAM when quantized.
A €0.02 Bank Transfer Can Turn a Banking AI Agent Into a Phishing Machine
Blue41's proof-of-concept against Bunq's AI assistant shows a single transaction description field can hijack the bank's own AI into delivering personalized phishing attacks — no malware, no device access needed.
Dario Amodei Publishes Policy Manifesto: Frontier Testing Law, Job Displacement Fund, and a Warning on Bio Risk
Anthropic's CEO argues AI policy is a year behind the exponential, names five areas requiring new law, and backs it with a legislative proposal on frontier model testing and a funded framework for job displacement.
Meta Built Six 125,000-Square-Foot Tents in Ohio to Cut Data Center Construction Time in Half
Meta has erected six rapid deployment structures at its New Albany, Ohio campus, each 125,000 sq ft and powered by 200MW of modular gas turbines. The tent strategy, borrowed from Tesla and xAI, is intended to halve the time from permit to live compute.
Opus 4.8 Thinking Enters Agent Arena at Rank 2 — Grok Build Lands at 17, xAI Sweeps the Bottom Six
Claude Opus 4.8 (Thinking) entered Agent Arena on June 9 and immediately ranked second with a 9.06% composite score, 0.05 points behind GPT-5.5 High. Without thinking mode, the same model drops to rank 8. All three xAI entries — Grok Build 0.1, Grok 4.3 High, and Grok 4.3 — rank 17, 19, and 22.
SpaceX IPO Demand Hits $250B, 4x Oversubscribed — xAI Has Lost $41.3B and Consumed 76% of Q1 Capex
SpaceX prices its IPO Thursday at $135 per share ($1.77T valuation) with $250 billion in demand — four times the $75B raise target. Inside the S-1: xAI's AI segment accumulated $41.3B in total losses, consumed 76% of all SpaceX capex in Q1 2026, and generated $3.2B in AI revenue against $6.4B in operating losses in 2025 alone.
Big Tech Will Spend $600B on AI in 2026 and Barely Bought Anyone Building It
PitchBook's Q2 2026 analysis puts combined AI infrastructure spending by Google, Microsoft, Amazon, and Meta at $600B for the year. M&A hit a decade low in 2024 and has barely recovered. The hyperscalers are building and backing as minority investors, not acquiring.
Germany Makes Google Liable for AI Search Errors as UK Lets Publishers Block Them
A German court ruled Google's AI Overview summaries are its own speech, making it liable for false claims. The UK CMA the same week ordered publishers given binding rights to opt content out of AI search features and training data entirely.
Claude Fable 5 Takes #1 on Artificial Analysis Intelligence Index at 64.9 — Anthropic Holds Both Top Spots
Fable 5 scores 64.9 on the 10-benchmark AA Intelligence Index, pushing Opus 4.8 (61.4) to #2 and GPT-5.5 to third. On Humanity's Last Exam, Fable 5 hits 53% — more than 7 points ahead of any other model evaluated.
OpenAI's Dreaming V3 Scores 82.8% on Recall -- and Lost the Feature Power Users Actually Used
Dreaming V3, rolled out June 4, eliminated ChatGPT's explicit saved-memories notebook in favour of automated background synthesis. Factual recall improved. User control over what gets remembered did not survive.